Home | History | Annotate | Download | only in AMDGPU 
      1 ; RUN: llc -march=amdgcn -mtriple=amdgcn-unknown-amdhsa -mattr=-promote-alloca,+max-private-element-size-16 -verify-machineinstrs < %s | FileCheck -check-prefix=ELT16 -check-prefix=HSA -check-prefix=HSA-ELT16 -check-prefix=ALL -check-prefix=HSA_ELTGE8 %s
      2 ; RUN: llc -march=amdgcn -mtriple=amdgcn-unknown-amdhsa -mattr=-promote-alloca,+max-private-element-size-8 -verify-machineinstrs < %s | FileCheck -check-prefix=ELT8 -check-prefix=HSA -check-prefix=HSA-ELT8 -check-prefix=ALL -check-prefix=HSA-ELTGE8 %s
      3 ; RUN: llc -march=amdgcn -mtriple=amdgcn-unknown-amdhsa -mattr=-promote-alloca,+max-private-element-size-4 -verify-machineinstrs < %s | FileCheck -check-prefix=ELT4 -check-prefix=HSA -check-prefix=HSA-ELT4 -check-prefix=ALL %s
      4 
      5 
      6 ; ALL-LABEL: {{^}}private_elt_size_v4i32:
      7 
      8 ; HSA-ELT16: private_element_size = 3
      9 ; HSA-ELT8: private_element_size = 2
     10 ; HSA-ELT4: private_element_size = 1
     11 
     12 
     13 ; HSA-ELT16-DAG: buffer_store_dwordx4 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen{{$}}
     14 ; HSA-ELT16-DAG: buffer_store_dwordx4 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen offset:16
     15 ; HSA-ELT16-DAG: buffer_load_dwordx4 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen{{$}}
     16 
     17 ; HSA-ELT8-DAG: buffer_store_dwordx2 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen{{$}}
     18 ; HSA-ELT8-DAG: buffer_store_dwordx2 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen offset:8
     19 ; HSA-ELT8-DAG: buffer_store_dwordx2 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen offset:16
     20 ; HSA-ELT8-DAG: buffer_store_dwordx2 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen offset:24
     21 
     22 ; HSA-ELT8: buffer_load_dwordx2 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen
     23 ; HSA-ELT8: buffer_load_dwordx2 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen
     24 
     25 
     26 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen{{$}}
     27 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:4{{$}}
     28 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:8{{$}}
     29 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:12{{$}}
     30 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:16{{$}}
     31 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:20{{$}}
     32 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:24{{$}}
     33 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:28{{$}}
     34 
     35 ; HSA-ELT4: buffer_load_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen{{$}}
     36 ; HSA-ELT4: buffer_load_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:4{{$}}
     37 ; HSA-ELT4: buffer_load_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:8{{$}}
     38 ; HSA-ELT4: buffer_load_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:12{{$}}
     39 define void @private_elt_size_v4i32(<4 x i32> addrspace(1)* %out, i32 addrspace(1)* %index.array) #0 {
     40 entry:
     41   %tid = call i32 @llvm.amdgcn.workitem.id.x()
     42   %idxprom = sext i32 %tid to i64
     43   %gep.index = getelementptr inbounds i32, i32 addrspace(1)* %index.array, i64 %idxprom
     44   %index.load = load i32, i32 addrspace(1)* %gep.index
     45   %index = and i32 %index.load, 2
     46   %alloca = alloca [2 x <4 x i32>], align 16
     47   %gep0 = getelementptr inbounds [2 x <4 x i32>], [2 x <4 x i32>]* %alloca, i32 0, i32 0
     48   %gep1 = getelementptr inbounds [2 x <4 x i32>], [2 x <4 x i32>]* %alloca, i32 0, i32 1
     49   store <4 x i32> zeroinitializer, <4 x i32>* %gep0
     50   store <4 x i32> <i32 1, i32 2, i32 3, i32 4>, <4 x i32>* %gep1
     51   %gep2 = getelementptr inbounds [2 x <4 x i32>], [2 x <4 x i32>]* %alloca, i32 0, i32 %index
     52   %load = load <4 x i32>, <4 x i32>* %gep2
     53   store <4 x i32> %load, <4 x i32> addrspace(1)* %out
     54   ret void
     55 }
     56 
     57 ; ALL-LABEL: {{^}}private_elt_size_v8i32:
     58 ; HSA-ELT16: private_element_size = 3
     59 ; HSA-ELT8: private_element_size = 2
     60 ; HSA-ELT4: private_element_size = 1
     61 
     62 ; HSA-ELT16-DAG: buffer_store_dwordx4 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen{{$}}
     63 ; HSA-ELT16-DAG: buffer_store_dwordx4 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen offset:16
     64 ; HSA-ELT16-DAG: buffer_store_dwordx4 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen offset:32
     65 ; HSA-ELT16-DAG: buffer_store_dwordx4 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen offset:48
     66 
     67 ; HSA-ELT16-DAG: buffer_load_dwordx4 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen{{$}}
     68 ; HSA-ELT16-DAG: buffer_load_dwordx4 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen{{$}}
     69 
     70 
     71 ; HSA-ELT8-DAG: buffer_store_dwordx2 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen{{$}}
     72 ; HSA-ELT8-DAG: buffer_store_dwordx2 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen offset:8
     73 ; HSA-ELT8-DAG: buffer_store_dwordx2 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen offset:16
     74 ; HSA-ELT8-DAG: buffer_store_dwordx2 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen offset:24
     75 ; HSA-ELT8-DAG: buffer_store_dwordx2 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen offset:32
     76 ; HSA-ELT8-DAG: buffer_store_dwordx2 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen offset:40
     77 ; HSA-ELT8-DAG: buffer_store_dwordx2 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen offset:48
     78 ; HSA-ELT8-DAG: buffer_store_dwordx2 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen offset:56
     79 
     80 ; HSA-ELT8: buffer_load_dwordx2 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen
     81 ; HSA-ELT8: buffer_load_dwordx2 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen
     82 
     83 
     84 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen{{$}}
     85 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:4{{$}}
     86 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:8{{$}}
     87 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:12{{$}}
     88 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:16{{$}}
     89 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:20{{$}}
     90 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:24{{$}}
     91 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:28{{$}}
     92 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:32{{$}}
     93 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:36{{$}}
     94 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:40{{$}}
     95 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:44{{$}}
     96 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:48{{$}}
     97 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:52{{$}}
     98 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:56{{$}}
     99 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:60{{$}}
    100 
    101 ; HSA-ELT4-DAG: buffer_load_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen{{$}}
    102 ; HSA-ELT4-DAG: buffer_load_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:4{{$}}
    103 ; HSA-ELT4-DAG: buffer_load_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:8{{$}}
    104 ; HSA-ELT4-DAG: buffer_load_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:12{{$}}
    105 ; HSA-ELT4-DAG: buffer_load_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:16{{$}}
    106 ; HSA-ELT4-DAG: buffer_load_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:20{{$}}
    107 ; HSA-ELT4-DAG: buffer_load_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:24{{$}}
    108 ; HSA-ELT4-DAG: buffer_load_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:28{{$}}
    109 define void @private_elt_size_v8i32(<8 x i32> addrspace(1)* %out, i32 addrspace(1)* %index.array) #0 {
    110 entry:
    111   %tid = call i32 @llvm.amdgcn.workitem.id.x()
    112   %idxprom = sext i32 %tid to i64
    113   %gep.index = getelementptr inbounds i32, i32 addrspace(1)* %index.array, i64 %idxprom
    114   %index.load = load i32, i32 addrspace(1)* %gep.index
    115   %index = and i32 %index.load, 2
    116   %alloca = alloca [2 x <8 x i32>], align 16
    117   %gep0 = getelementptr inbounds [2 x <8 x i32>], [2 x <8 x i32>]* %alloca, i32 0, i32 0
    118   %gep1 = getelementptr inbounds [2 x <8 x i32>], [2 x <8 x i32>]* %alloca, i32 0, i32 1
    119   store <8 x i32> zeroinitializer, <8 x i32>* %gep0
    120   store <8 x i32> <i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8>, <8 x i32>* %gep1
    121   %gep2 = getelementptr inbounds [2 x <8 x i32>], [2 x <8 x i32>]* %alloca, i32 0, i32 %index
    122   %load = load <8 x i32>, <8 x i32>* %gep2
    123   store <8 x i32> %load, <8 x i32> addrspace(1)* %out
    124   ret void
    125 }
    126 
    127 
    128 ; ALL-LABEL: {{^}}private_elt_size_i64:
    129 ; HSA-ELT16: private_element_size = 3
    130 ; HSA-ELT8: private_element_size = 2
    131 ; HSA-ELT4: private_element_size = 1
    132 
    133 ; HSA-ELTGE8-DAG: buffer_store_dwordx2 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen{{$}}
    134 ; HSA-ELTGE8-DAG: buffer_store_dwordx2 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen offset:8
    135 
    136 ; HSA-ELTGE8: buffer_load_dwordx2 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen
    137 
    138 
    139 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen{{$}}
    140 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:4{{$}}
    141 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:8{{$}}
    142 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:12{{$}}
    143 
    144 ; HSA-ELT4: buffer_load_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen{{$}}
    145 ; HSA-ELT4: buffer_load_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:4{{$}}
    146 define void @private_elt_size_i64(i64 addrspace(1)* %out, i32 addrspace(1)* %index.array) #0 {
    147 entry:
    148   %tid = call i32 @llvm.amdgcn.workitem.id.x()
    149   %idxprom = sext i32 %tid to i64
    150   %gep.index = getelementptr inbounds i32, i32 addrspace(1)* %index.array, i64 %idxprom
    151   %index.load = load i32, i32 addrspace(1)* %gep.index
    152   %index = and i32 %index.load, 2
    153   %alloca = alloca [2 x i64], align 16
    154   %gep0 = getelementptr inbounds [2 x i64], [2 x i64]* %alloca, i32 0, i32 0
    155   %gep1 = getelementptr inbounds [2 x i64], [2 x i64]* %alloca, i32 0, i32 1
    156   store i64 0, i64* %gep0
    157   store i64 34359738602, i64* %gep1
    158   %gep2 = getelementptr inbounds [2 x i64], [2 x i64]* %alloca, i32 0, i32 %index
    159   %load = load i64, i64* %gep2
    160   store i64 %load, i64 addrspace(1)* %out
    161   ret void
    162 }
    163 
    164 ; ALL-LABEL: {{^}}private_elt_size_f64:
    165 ; HSA-ELT16: private_element_size = 3
    166 ; HSA-ELT8: private_element_size = 2
    167 ; HSA-ELT4: private_element_size = 1
    168 
    169 ; HSA-ELTGE8-DAG: buffer_store_dwordx2 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen{{$}}
    170 ; HSA-ELTGE8-DAG: buffer_store_dwordx2 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen offset:8
    171 
    172 ; HSA-ELTGE8: buffer_load_dwordx2 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen
    173 
    174 
    175 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen{{$}}
    176 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:4{{$}}
    177 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:8{{$}}
    178 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:12{{$}}
    179 
    180 ; HSA-ELT4: buffer_load_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen{{$}}
    181 ; HSA-ELT4: buffer_load_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:4{{$}}
    182 define void @private_elt_size_f64(double addrspace(1)* %out, i32 addrspace(1)* %index.array) #0 {
    183 entry:
    184   %tid = call i32 @llvm.amdgcn.workitem.id.x()
    185   %idxprom = sext i32 %tid to i64
    186   %gep.index = getelementptr inbounds i32, i32 addrspace(1)* %index.array, i64 %idxprom
    187   %index.load = load i32, i32 addrspace(1)* %gep.index
    188   %index = and i32 %index.load, 2
    189   %alloca = alloca [2 x double], align 16
    190   %gep0 = getelementptr inbounds [2 x double], [2 x double]* %alloca, i32 0, i32 0
    191   %gep1 = getelementptr inbounds [2 x double], [2 x double]* %alloca, i32 0, i32 1
    192   store double 0.0, double* %gep0
    193   store double 4.0, double* %gep1
    194   %gep2 = getelementptr inbounds [2 x double], [2 x double]* %alloca, i32 0, i32 %index
    195   %load = load double, double* %gep2
    196   store double %load, double addrspace(1)* %out
    197   ret void
    198 }
    199 
    200 ; ALL-LABEL: {{^}}private_elt_size_v2i64:
    201 ; HSA-ELT16: private_element_size = 3
    202 ; HSA-ELT8: private_element_size = 2
    203 ; HSA-ELT4: private_element_size = 1
    204 
    205 ; HSA-ELT16-DAG: buffer_store_dwordx4 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen{{$}}
    206 ; HSA-ELT16-DAG: buffer_store_dwordx4 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen offset:16
    207 ; HSA-ELT16-DAG: buffer_load_dwordx4 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen{{$}}
    208 
    209 ; HSA-ELT8-DAG: buffer_store_dwordx2 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen{{$}}
    210 ; HSA-ELT8-DAG: buffer_store_dwordx2 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen offset:8
    211 ; HSA-ELT8-DAG: buffer_store_dwordx2 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen offset:16
    212 ; HSA-ELT8-DAG: buffer_store_dwordx2 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen offset:24
    213 
    214 ; HSA-ELT8: buffer_load_dwordx2 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen
    215 ; HSA-ELT8: buffer_load_dwordx2 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, s[0:3], s9 offen
    216 
    217 
    218 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen{{$}}
    219 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:4{{$}}
    220 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:8{{$}}
    221 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:12{{$}}
    222 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:16{{$}}
    223 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:20{{$}}
    224 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:24{{$}}
    225 ; HSA-ELT4-DAG: buffer_store_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:28{{$}}
    226 
    227 ; HSA-ELT4: buffer_load_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen{{$}}
    228 ; HSA-ELT4: buffer_load_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:4{{$}}
    229 ; HSA-ELT4: buffer_load_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:8{{$}}
    230 ; HSA-ELT4: buffer_load_dword {{v[0-9]+}}, v{{[0-9]+}}, s[0:3], s9 offen offset:12{{$}}
    231 define void @private_elt_size_v2i64(<2 x i64> addrspace(1)* %out, i32 addrspace(1)* %index.array) #0 {
    232 entry:
    233   %tid = call i32 @llvm.amdgcn.workitem.id.x()
    234   %idxprom = sext i32 %tid to i64
    235   %gep.index = getelementptr inbounds i32, i32 addrspace(1)* %index.array, i64 %idxprom
    236   %index.load = load i32, i32 addrspace(1)* %gep.index
    237   %index = and i32 %index.load, 2
    238   %alloca = alloca [2 x <2 x i64>], align 16
    239   %gep0 = getelementptr inbounds [2 x <2 x i64>], [2 x <2 x i64>]* %alloca, i32 0, i32 0
    240   %gep1 = getelementptr inbounds [2 x <2 x i64>], [2 x <2 x i64>]* %alloca, i32 0, i32 1
    241   store <2 x i64> zeroinitializer, <2 x i64>* %gep0
    242   store <2 x i64> <i64 1, i64 2>, <2 x i64>* %gep1
    243   %gep2 = getelementptr inbounds [2 x <2 x i64>], [2 x <2 x i64>]* %alloca, i32 0, i32 %index
    244   %load = load <2 x i64>, <2 x i64>* %gep2
    245   store <2 x i64> %load, <2 x i64> addrspace(1)* %out
    246   ret void
    247 }
    248 
    249 declare i32 @llvm.amdgcn.workitem.id.x() #1
    250 
    251 attributes #0 = { nounwind }
    252 attributes #1 = { nounwind readnone }
    253